Merge branch 'main' into toupstream/sub_scalar

Author: SaoirseARM

CMakeLists.txt

@@ -48,8 +48,6 @@
 cmake_minimum_required(VERSION 3.24)
 project(executorch)
 
-# MARK: - Start EXECUTORCH_H12025_BUILD_MIGRATION
-
 include(${PROJECT_SOURCE_DIR}/tools/cmake/common/preset.cmake)
 include(${PROJECT_SOURCE_DIR}/tools/cmake/Utils.cmake)
 include(CMakeDependentOption)
@@ -82,6 +80,7 @@ announce_configured_options(BUCK2)
 
 announce_configured_options(CMAKE_CXX_COMPILER_ID)
 announce_configured_options(CMAKE_TOOLCHAIN_FILE)
+announce_configured_options(BUILD_TESTING)
 
 load_build_preset()
 include(${PROJECT_SOURCE_DIR}/tools/cmake/preset/default.cmake)
@@ -97,11 +96,6 @@ else()
 endif()
 announce_configured_options(CCACHE_PROGRAM)
 
-# Print all the configs that were called with announce_configured_options.
-print_configured_options()
-
-# MARK: - End EXECUTORCH_H12025_BUILD_MIGRATION
-
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 # Setup RPATH. See
@@ -750,3 +744,6 @@ if(EXECUTORCH_BUILD_ANDROID_JNI)
 endif()
 
 include(Test.cmake)
+
+# Print all the configs that were called with announce_configured_options.
+print_configured_options()

CMakePresets.json

@@ -8,7 +8,7 @@
     },
     {
       "name": "macos",
-      "displayName": "Build everything buildable on macOS",
+      "displayName": "Build ExecuTorch for macOS",
       "inherits": ["common"],
       "generator": "Xcode",
       "cacheVariables": {
@@ -25,7 +25,7 @@
     },
     {
       "name": "ios",
-      "displayName": "Build everything buildable on iOS",
+      "displayName": "Build ExecuTorch for iOS",
       "inherits": ["common"],
       "generator": "Xcode",
       "cacheVariables": {
@@ -42,7 +42,7 @@
     },
     {
       "name": "ios-simulator",
-      "displayName": "Build everything buildable on iOS simulator",
+      "displayName": "Build ExecuTorch for iOS Simulator",
       "inherits": ["common"],
       "generator": "Xcode",
       "cacheVariables": {
@@ -59,7 +59,7 @@
     },
     {
       "name": "linux",
-      "displayName": "Build everything buildable on Linux",
+      "displayName": "Build ExecuTorch for Linux",
       "inherits": ["common"],
       "cacheVariables": {
         "CMAKE_SYSTEM_NAME": "Linux",
@@ -88,29 +88,21 @@
     {
         "name": "llm",
         "displayName": "Build LLM libraries",
-        "inherits": [
-            "common"
-        ],
+        "inherits": ["common"],
         "cacheVariables": {
             "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/llm.cmake",
             "CMAKE_OSX_DEPLOYMENT_TARGET": "12.0"
         },
         "condition": {
             "type": "inList",
             "string": "${hostSystemName}",
-            "list": [
-                "Darwin",
-                "Linux",
-                "Windows"
-            ]
+            "list": ["Darwin", "Linux", "Windows"]
         }
     },
     {
         "name": "zephyr",
-        "displayName": "Build everything buildable on Zephyr RTOS",
-        "inherits": [
-            "common"
-        ],
+        "displayName": "Build ExecuTorch for Zephyr RTOS",
+        "inherits": ["common"],
         "cacheVariables": {
             "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/zephyr.cmake",
             "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/examples/zephyr/x86_64-linux-arm-zephyr-eabi-gcc.cmake"

backends/cadence/aot/tests/test_fusion_ops_passes.py

@@ -12,7 +12,6 @@
 
 import executorch.backends.cadence.aot.ops_registrations  # noqa
 import torch
-from executorch.backends.cadence.aot import compiler
 from executorch.backends.cadence.aot.fuse_ops import (
     FuseCascadedTransposeOrPermuteOps,
     FuseCascadedViewOps,
@@ -30,7 +29,6 @@
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.dialects.edge._ops import EdgeOpOverload
 from executorch.exir.pass_base import PassResult, ProxyValue
-from torch import nn
 
 
 class TestFusionPassesBase(unittest.TestCase):
@@ -178,43 +176,6 @@ def test_keep_mm_add_with_multiple_users(self) -> None:
         self.assertEqual(count_node(converted_graph, exir_ops.edge.aten.mm.default), 1)
         self.assertEqual(count_node(converted_graph, exir_ops.edge.aten.add.Tensor), 3)
 
-    # TODO(matthiascremon) -> None: enable that pass with new flow
-    @torch.no_grad()
-    @unittest.expectedFailure
-    def test_legacy_conv_bn_fusion(self) -> None:
-        class ModelConvBN(torch.nn.Module):
-            def __init__(
-                self, in_features: int, out_features: int, kernel_size: int
-            ) -> None:
-                super().__init__()
-                self.conv1d = nn.Conv1d(in_features, out_features, kernel_size)
-                self.bn = nn.BatchNorm1d(out_features)
-
-            def forward(self, x: torch.Tensor) -> torch.Tensor:
-                y = self.conv1d(x)
-                return self.bn(y)
-
-        model = ModelConvBN(64, 1, 2)
-        x = torch.randn(1, 64, 4)
-
-        graph_module = (
-            compiler.export_to_executorch_gen_etrecord(model.eval(), (x,))
-            .exported_program()
-            .graph_module
-        )
-        # Assert that after running the fusion passes, batchnorm was fused with conv1d
-        self.assertEqual(
-            count_node(graph_module, torch.ops.aten.linear.out)
-            + count_node(graph_module, torch.ops.cadence.convolution.out),
-            1,
-        )
-        self.assertEqual(
-            count_node(
-                graph_module, torch.ops.aten._native_batch_norm_legit_no_training.out
-            ),
-            0,
-        )
-
     def test_permute_transpose_fusion(self) -> None:
         builder = GraphBuilder()
         x = builder.placeholder("x", torch.randn(3, 1, 3, 1, 4, dtype=torch.float32))

backends/nxp/backend/edge_program_converter.py

@@ -31,6 +31,7 @@
     exir_ops.edge.aten.mm.default: MMConverter,  # noqa F405
     exir_ops.edge.aten.permute_copy.default: PermuteCopyConverter,  # noqa F405
     exir_ops.edge.aten.relu.default: ReLUConverter,  # noqa F405
+    exir_ops.edge.aten.hardtanh.default: HardTanhConverter,  # noqa F405
     exir_ops.edge.aten._softmax.default: SoftmaxConverter,  # noqa F405
     exir_ops.edge.aten.view_copy.default: ViewCopyConverter,  # noqa F405
 }

backends/nxp/backend/ir/converter/node_converters/ops_converters/__init__.py

@@ -10,6 +10,9 @@
 from executorch.backends.nxp.backend.ir.converter.node_converters.ops_converters.convolution_converter import (
     ConvolutionConverter,
 )
+from executorch.backends.nxp.backend.ir.converter.node_converters.ops_converters.hardtanh_converter import (
+    HardTanhConverter,
+)
 from executorch.backends.nxp.backend.ir.converter.node_converters.ops_converters.max_pool_2d_converter import (
     MaxPool2dConverter,
 )
@@ -48,4 +51,5 @@
     "ReLUConverter",
     "MaxPool2dConverter",
     "AvgPool2dConverter",
+    "HardTanhConverter",
 ]

backends/nxp/backend/ir/converter/node_converters/ops_converters/hardtanh_converter.py

@@ -0,0 +1,47 @@
+# Copyright (c) 2025 NXP
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from executorch.backends.nxp.backend.ir.converter.node_converter import (
+    NodeConverter,
+    Target,
+)
+from executorch.backends.nxp.backend.ir.lib.tflite.BuiltinOperator import (
+    BuiltinOperator,
+)
+from torch.fx import Node
+from torch.nn import Parameter
+
+
+class HardTanhConverter(NodeConverter):
+    supported_targets = [Target.RT700]
+
+    # Maps possible input parameters of HardTanh to equivalent ReLU-based operators supported by TFLite.
+    supported_modes_map = {
+        (0.0, 6.0): BuiltinOperator.RELU6,
+        (-1.0, 1.0): BuiltinOperator.RELU_N1_TO_1,
+        (0.0, 1.0): BuiltinOperator.RELU_0_TO_1,
+        (0.0, float("inf")): BuiltinOperator.RELU,
+    }
+
+    @staticmethod
+    def _is_supported_in_IR(
+        node: Node, parameters_mapping: dict[str, Parameter]
+    ) -> bool:
+        _, min_value, max_value = node.args
+        return (min_value, max_value) in HardTanhConverter.supported_modes_map.keys()
+
+    def convert(self, node: Node):
+        """Convert 'aten::hardtanh' to it's supported ReLU equivalent."""
+        self.assert_convertible(node)
+
+        t_op = self._create_tflite_op_with_io_tensors(node)
+
+        _, min_value, max_value = node.args
+
+        op = self.supported_modes_map[(min_value, max_value)]
+        t_op.opcode_index = self.builder.op_code_index_for_op_type(op)
+
+        self.builder.append_operators([t_op])

backends/nxp/neutron_partitioner.py

@@ -195,6 +195,7 @@ def tag_qdq_clusters(self, nodes: List[torch.fx.Node]):
     exir_ops.edge.aten.max_pool2d_with_indices.default: MaxPool2dConverter,  # noqa F405
     exir_ops.edge.aten.mm.default: MMConverter,  # noqa F405
     exir_ops.edge.aten.relu.default: ReLUConverter,  # noqa F405
+    exir_ops.edge.aten.hardtanh.default: HardTanhConverter,  # noqa F405
     exir_ops.edge.aten._softmax.default: SoftmaxConverter,  # noqa F405
     exir_ops.edge.aten.view_copy.default: ViewCopyConverter,  # noqa F405
 }

backends/nxp/nxp_backend.py

@@ -174,7 +174,8 @@ def preprocess(
             # Otherwise, we get violation that this op is not part of ATen Core ops.
             edge_program._verifiers = [
                 EXIREdgeDialectVerifier(
-                    class_only=True, exception_list=[torch.ops.aten.max_pool2d.default]
+                    class_only=True,
+                    core_aten_ops_exception_list=[torch.ops.aten.max_pool2d.default],
                 )
             ]
 

backends/nxp/quantizer/neutron_quantizer.py

@@ -15,6 +15,8 @@
     AvgPoolPattern,
     Conv1dPattern,
     Conv2dPattern,
+    HardTanhInPlacePattern,
+    HardTanhPattern,
     LinearPattern,
     MaxPoolPattern,
     PadPattern,
@@ -199,6 +201,8 @@ def __init__(self):
                 NeutronAtenQuantizer(PermutePattern(), static_qconfig),
                 NeutronAtenQuantizer(PadPattern(), static_qconfig),
                 NeutronAtenQuantizer(ReluPattern(), static_qconfig),
+                NeutronAtenQuantizer(HardTanhPattern(), static_qconfig),
+                NeutronAtenQuantizer(HardTanhInPlacePattern(), static_qconfig),
                 NeutronAtenQuantizer(ReluInPlacePattern(), static_qconfig),
                 NeutronAtenQuantizer(AvgPoolPattern(), static_qconfig),
                 NeutronAtenQuantizer(ViewPattern(), static_qconfig),

backends/nxp/quantizer/patterns.py

@@ -216,6 +216,26 @@ def get_anchors(
         )
 
 
+class HardTanhPattern(SharedSpecPattern):
+    """
+    Quantizer for HardTanh operator. Shared quantization spec is selected, as activation functions usually follows
+    computation layer.
+    """
+
+    def partition_types(self):
+        return [torch.ops.aten.hardtanh.default]
+
+
+class HardTanhInPlacePattern(SharedSpecPattern):
+    """
+    Quantizer for HardTanh operator with param inplace=True. Shared quantization spec is selected, as activation
+    functions usually follows computation layer.
+    """
+
+    def partition_types(self):
+        return [torch.ops.aten.hardtanh_.default]
+
+
 class LinearPattern(QuantizationPattern):
     def partition_types(self) -> List[OpOverload]:
         return [torch.ops.aten.linear.default]